Quantifying urban solar potential losses from rooftop superstructures via aerial imagery and Convolutional Neural Networks

Accurately estimating the available rooftop area for solar power production systems is crucial for precise assessments of urban solar potential. A primary challenge in this estimation is the presence of rooftop superstructures (e.g., chimneys, heating, ventilation, and air conditioning systems) that hinder installations. In this research, we develop a deep learning-based approach to identify rooftop obstructions using aerial imagery of approximately 35,000 buildings within the Canton of Geneva. Automatic labeling techniques are used to generate training and testing datasets. A semantic segmentation model, based on a U-Net architecture with a ResNet-152 backbone, detects superstructures, achieving an Intersection over Union (IoU) of 0.40. Subsequently, a binary classification of the roof segments into “free” or “obstructed” categories is performed based on the presence of superstructures, achieving an accuracy of 85%. Cross-validation on 1500 manually labeled buildings in Switzerland yields an accuracy of 91%. The results show that excluding roof segments with superstructures reduces the estimated solar potential by an average of 43%, ranging from 31% in suburban areas to 51% in urban centers. This significant reduction, along with the variability across different urban zones, highlights the importance of integrating obstruction detection into urban solar potential estimations through scalable and automated solutions.